The only relevant thing in the link is the improvement in benchmarks, which is real but doesn't involve the kind of "auto-improvement" you seem to think exists.

the devs don't even understand anymore the script iimproved by the AI itself

AI doesn't write some magical code that nobody understands. Understanding any code takes effort, doubly so if the code is sloppy and effort goes against the whole philosophy of offloading mental tasks to ai.

A human made sandbox is even currently a joke for some AIs, breaking out in few minutes as already ocurred.

They told an LLM made to find vulnerabilities in software to try escaping its sandbox, and it found the vulnerabilities to do so. That's an indication that there're still vulnerabilities to be discovered and fixed, not that "ai can't be contained".

Quantum computers with AI

Ai has nothing to do with quantum computers. If one day quantum computers advance enough to be useful for actual computation, running ai on them would still require a completely different architecture from the llms of today.

The irl problem with 'ai' isn't a skynet scenario, it's the resources it consumes, the reduction in quality when it is used to make things and the lives of the people it upheavals.